library(dplyr)
library(here)
library(ggplot2)
library(tidyr)
library(ggthemes)
library(doBy)
library(reshape)
library(plotly)
library(GGally)
library(caret)
library(e1071)
library(class)
library(usmap)
First, we create some custom theme to ensure homogoneity throughout the presentation deck.
# set global theme
# custom theme
bg_color = "#0B0C10"
bar_color = "#66FCF1"
text_main = "#FFFFF"
text_ticks = "#CFC6C7"
axis_lines = #453a3b"
cust_theme <- theme(plot.title = element_text(color = 'white', vjust=0.5, hjust=0.5),
plot.background = element_rect(fill = bg_color),
panel.background = element_rect(fill = bg_color),
axis.text.x=element_text(angle=90,hjust=1),
axis.text = element_text(colour = text_ticks),
panel.grid.major.x = element_blank(),
panel.grid.minor.x = element_blank(),
panel.grid.major.y = element_line(colour = "#453a3b", linetype = 1, size = 0.25),
panel.grid.minor.y = element_blank(),
axis.title = element_text(colour = "white"))
# overwrite default theme with custom
theme_set(theme_foundation() + cust_theme)
Then, we created a relative file path from the “here” library for better reproducibility. After setting our working directory, we imported the files into dataframes for better transformation capabilities.
# set relative brewery file path as variable
beers_csv <- here("project_files", "Beers.csv")
breweries_csv <- here("project_files", "Breweries.csv")
# initialize dataframes
beers <- data.frame(read.csv(beers_csv))
breweries <- data.frame(read.csv(breweries_csv))
How many breweries are present in each state? In order to investigate the number of breweries in our data set, we group everything by state and then perform a tally function in order to count total breweries. Then we plot the results as a bar chart using ggplot (and plotly).
# Question 1 - How many breweries are present in each state?
per_state <- breweries %>% group_by(State) %>% tally(name="breweries_per_state") %>% arrange(desc(breweries_per_state))
fig1 <- per_state %>% ggplot(aes(x=reorder(State, breweries_per_state), y=breweries_per_state)) + geom_bar(stat="identity", fill=bar_color) + ggtitle("Total Breweries per State") + ylab("Total Breweries") + xlab("State")
ggplotly(fig1)
# Part 2: Joining beer and brewery datasets. Can we combine the data? What’s it look like? In order to join the two dataframes, we called the “merge” function and joined the two dataframes on common keys - brewery_id and brew_id. We made sure to rename the columns for easier reuse and then displayed the results using an rbind function to combine the heads and tails of the resultant dataframe.
# Question 2: Merge Beer data with Breweries data. Print head(6) and tail(6)
# Note: left join breweries on beers because 1-many relationship
main <- merge(beers, breweries, by.x="Brewery_id", by.y="Brew_ID")
# clean column names for good housekeeping
main_cols <- c("brewery_id", "beer_name", "beer_id", "abv", "ibu", "beer_style", "serving_ounces", "brewery_name", "city", "state")
names(main) <- main_cols
# print head and tail of resultant data set
print(rbind(head(main, 6), tail(main, 6)))
## brewery_id beer_name beer_id abv ibu
## 1 1 Get Together 2692 0.045 50
## 2 1 Maggie's Leap 2691 0.049 26
## 3 1 Wall's End 2690 0.048 19
## 4 1 Pumpion 2689 0.060 38
## 5 1 Stronghold 2688 0.060 25
## 6 1 Parapet ESB 2687 0.056 47
## 2405 556 Pilsner Ukiah 98 0.055 NA
## 2406 557 Heinnieweisse Weissebier 52 0.049 NA
## 2407 557 Snapperhead IPA 51 0.068 NA
## 2408 557 Moo Thunder Stout 50 0.049 NA
## 2409 557 Porkslap Pale Ale 49 0.043 NA
## 2410 558 Urban Wilderness Pale Ale 30 0.049 NA
## beer_style serving_ounces
## 1 American IPA 16
## 2 Milk / Sweet Stout 16
## 3 English Brown Ale 16
## 4 Pumpkin Ale 16
## 5 American Porter 16
## 6 Extra Special / Strong Bitter (ESB) 16
## 2405 German Pilsener 12
## 2406 Hefeweizen 12
## 2407 American IPA 12
## 2408 Milk / Sweet Stout 12
## 2409 American Pale Ale (APA) 12
## 2410 English Pale Ale 12
## brewery_name city state
## 1 NorthGate Brewing Minneapolis MN
## 2 NorthGate Brewing Minneapolis MN
## 3 NorthGate Brewing Minneapolis MN
## 4 NorthGate Brewing Minneapolis MN
## 5 NorthGate Brewing Minneapolis MN
## 6 NorthGate Brewing Minneapolis MN
## 2405 Ukiah Brewing Company Ukiah CA
## 2406 Butternuts Beer and Ale Garrattsville NY
## 2407 Butternuts Beer and Ale Garrattsville NY
## 2408 Butternuts Beer and Ale Garrattsville NY
## 2409 Butternuts Beer and Ale Garrattsville NY
## 2410 Sleeping Lady Brewing Company Anchorage AK
How can we address the “NA” values? In order to perform more detailed analysis, it’s important to address the null values. In order to get a better idea of missing data, we filtered the dataframe to only null values using “is.na” and then took the count of each column using “colSums”. You can see in the print statement below that the most influencial null value is ibu. Almost half of the ibu values are missing! In order to “massage” this into workable data, we made an assumption that the median values for each style should be fairly close to the actual values. This is due to the fact that each beer style has ranges on ibu, therefore, we know a good range of potential values which the actual could lie in.
# Part 3: Address NA values in each column
# find the NA count per column to decide next steps...you will see some missing abv and A LOT of missing ibu
print(colSums(is.na(main)))
## brewery_id beer_name beer_id abv ibu
## 0 0 0 62 1005
## beer_style serving_ounces brewery_name city state
## 0 0 0 0 0
# Preserve a version of merged dataset by removing all NAs
main_clean<-na.omit(main)
# based on evidence, will use median abv and ibu per beer style to fill na
main <- main %>% group_by(beer_style) %>% mutate(ibu_corr = ifelse(is.na(ibu), median(ibu, na.rm = TRUE), ibu), abv_corr = ifelse(is.na(abv), median(abv, na.rm = TRUE), abv))
# let's see how we did...you will see all abv "corrected" and ibu had over 950 values "corrected"
print(colSums(is.na(main)))
## brewery_id beer_name beer_id abv ibu
## 0 0 0 62 1005
## beer_style serving_ounces brewery_name city state
## 0 0 0 0 0
## ibu_corr abv_corr
## 52 0
# Export the no-NAs file to a new csv file
write.csv(main_clean,"./Brewery_and_Beer_Clean.csv", row.names = FALSE)
After checking for null values, we then compute the median ABV and IBU for the dataset to give a representation of the central values to expect among all beers. We accomplish this using a grouping statement and then dplyr’s summarise statement in order to perform a median aggregation on each column.
# Part 4: Compute Median ABV and IBU and do bar plot
medians <- main_clean %>% group_by(state) %>% summarise(median_abv = median(abv), median_ibu = median(ibu))
#Bar_Chart_Plotter
# ibu bar plot
fig2 <- medians %>% ggplot() + geom_bar(aes(x=reorder(state, -median_ibu), y=median_ibu), stat="identity", fill=bar_color) + ggtitle("Median IBU per State") +ylab("Median IBU") + xlab("State")
# abv bar plot
fig3 <- medians %>% ggplot() + geom_bar(aes(x=reorder(state, -median_abv), y=median_abv*100), stat="identity", fill=bar_color) + ggtitle("Median ABV per State") +ylab("Median ABV") + xlab("State") + expand_limits(y=c(0, max(medians$median_abv+0.5)))
ggplotly(fig2)
ggplotly(fig3)
# Part 5: State with Max ABV and Max IBU. In order to get an idea of the maximum values in the dataset, we perform a simple lookup of maximum abv and ibu for the dataset. We do this by utilizing “which” to look up the index of the max value specified. From there, we filter our dataset to only include that particular index.
#Part5: The State with max ABV and max IBU
#The item with max ABV
beer_MaxAbv <- main_clean[which.max(main_clean$abv),]
print(paste0("The state has maximum alcoholic beer is:", beer_MaxAbv$state, " with ABV of ", beer_MaxAbv$abv))
## [1] "The state has maximum alcoholic beer is: KY with ABV of 0.125"
#The item with max IBU
beer_MaxIbu <- main_clean[which.max(main_clean$ibu),]
print(paste0("The state has maximum bitterness beer is:", beer_MaxIbu$state, " with IBU of ", beer_MaxIbu$ibu))
## [1] "The state has maximum bitterness beer is: OR with IBU of 138"
To get an idea of the overall statistical distribution of the abv data, we then used “summary” to return a statistical summary of the abv data and then create a histogram to show normality and spread.
#Part6: The summary statistics and distribution of the ABV
summary(main_clean$abv)
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.02700 0.05000 0.05700 0.05991 0.06800 0.12500
fig4 <- main_clean %>% ggplot() + geom_histogram(aes(x=abv*100), fill=bar_color) + ggtitle("Distribution of ABV") + xlab("Percent ABV")
ggplotly(fig4)
## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
# Part 7: Analysis of the relationship between ABV and IBU. Two of the most pricniple characteristics of beer are the ibu and abv measurements. What’s the relationship between these two? In order to test for this, we created a scatterplot of abv vs. ibu. You can see that the positive correlation between the two suggests that higher alcohol content is common among higher abv beers.
#Part7: Relationship between IBU and ABV?
#Scatterplot
fig5 <- main_clean %>% ggplot(aes(ibu, abv*100)) + geom_point(colour=bar_color, na.rm=TRUE)+geom_smooth(method=lm, se=FALSE,color="white", linetype="dashed", na.rm=TRUE) + ggtitle("ABV vs. IBU") + ylab("abv")
ggplotly(fig5)
# Part 8: KNN model creation based on IBU and ABV. Next, we will create an estimation model in order to classify beers (IPA vs. Non-IPA) based on their alchol and IBU levels. The first steps are preparing the data by splitting and classifying accordingly.
#Part8: KNN used for analysis IBU and ABV relationship of IPA and Ale(no IPA)
#Form Dataset with beer style IPA
main_IPA <- main_clean[grep("IPA",main_clean$beer_style),]
#Exclude the "Ale" part within IPA subset
main_IPA <- main_IPA[!grepl("Ale", main_IPA$beer_style),]
print(head(main_IPA))
## brewery_id beer_name beer_id abv ibu
## 1 1 Get Together 2692 0.045 50
## 7 2 Citra Ass Down 2686 0.080 68
## 15 2 Rico Sauvin 2678 0.076 68
## 18 2 Pile of Face 2675 0.060 65
## 25 4 Habitus (2014) 2668 0.080 100
## 26 4 Solis 2667 0.075 85
## beer_style serving_ounces brewery_name
## 1 American IPA 16 NorthGate Brewing
## 7 American Double / Imperial IPA 16 Against the Grain Brewery
## 15 American Double / Imperial IPA 16 Against the Grain Brewery
## 18 American IPA 16 Against the Grain Brewery
## 25 American Double / Imperial IPA 16 Mike Hess Brewing Company
## 26 American IPA 16 Mike Hess Brewing Company
## city state
## 1 Minneapolis MN
## 7 Louisville KY
## 15 Louisville KY
## 18 Louisville KY
## 25 San Diego CA
## 26 San Diego CA
#Form Dataset with beer style Ale
main_Ale <- main_clean[grep("Ale",main_clean$beer_style),]
#Exclude the "IPA" part within Ale subset
main_Ale <- main_Ale[!grepl("IPA", main_Ale$beer_style),]
print(head(main_Ale))
## brewery_id beer_name beer_id abv ibu beer_style
## 3 1 Wall's End 2690 0.048 19 English Brown Ale
## 4 1 Pumpion 2689 0.060 38 Pumpkin Ale
## 10 2 A Beer 2683 0.042 42 American Pale Ale (APA)
## 12 2 Flesh Gourd'n 2681 0.066 21 Pumpkin Ale
## 13 2 Sho'nuff 2680 0.040 13 Belgian Pale Ale
## 16 2 Coq de la Marche 2677 0.051 38 Saison / Farmhouse Ale
## serving_ounces brewery_name city state
## 3 16 NorthGate Brewing Minneapolis MN
## 4 16 NorthGate Brewing Minneapolis MN
## 10 16 Against the Grain Brewery Louisville KY
## 12 16 Against the Grain Brewery Louisville KY
## 13 16 Against the Grain Brewery Louisville KY
## 16 16 Against the Grain Brewery Louisville KY
Next, we will plot each seperate classification by their associated parameters in order to identify correlation between ABV and IBU for each group.
#Scatterplot of IPA and Ale datasets
fig6 <- main_IPA %>% ggplot(mapping = aes(ibu, abv)) + geom_point(colour = bar_color, na.rm=TRUE)+geom_smooth(method=lm, se=FALSE, na.rm=TRUE, colour="white") + ggtitle("IPA: ABV vs. IBU")
fig7 <- main_Ale %>% ggplot(mapping = aes(ibu, abv)) + geom_point(color = bar_color, na.rm=TRUE)+geom_smooth(method=lm, se=FALSE, na.rm=TRUE, colour="white") + ggtitle("Ale: ABV vs. IBU")
ggplotly(fig6)
ggplotly(fig7)
Now, we will rejoin the two groups together to continue to the modeling phase
#Select cols in IPA and Ale datasets and set group name
#Group numbers: Ale--1, IPA--2
test_Ale <- main_Ale %>% select(ibu,abv)
test_Ale$flag = "ALE"
test_IPA <- main_IPA %>% select(ibu,abv)
test_IPA$flag = "IPA"
test_4KNN <- rbind(test_Ale, test_IPA) #Combine two subsets into one for later classification
test_4KNN$flag <- as.factor(test_4KNN$flag)
You can see below how there are noticeable clusters being formed between IPA and non-IPA beers with the limit being somewhere around 50 IBU.
#Plot IPA and Ale datasets together in a single scatterplot
test_4KNN %>% ggplot(mapping = aes(ibu, abv,color=test_4KNN$flag)) + geom_point(na.rm=TRUE) + geom_smooth(method=lm, se=FALSE, na.rm=TRUE, linetype="dashed") + scale_colour_manual(name="Beer Style", values=c("ALE" = "#FF652F","IPA" ="#14A76C")) + ggtitle("ABV vs. IBU by Beer Style: KNN Test")
Next, we will split our dataset into train-split subsets. We have specified that 70% will be used as training set, and 30% will be used for testing.
#Divide dataset into train and test
set.seed(123)
trainIndex = sample(seq(1:937), 650)
trainBeers = test_4KNN[trainIndex,]
testBeers = test_4KNN[-trainIndex,]
#Train Data Visualization
trainBeers %>% ggplot(aes(x = abv,fill = flag)) + geom_histogram() + facet_grid(rows = vars(flag))
trainBeers %>% ggplot(aes(x = ibu,fill = flag)) + geom_histogram() + facet_grid(rows = vars(flag))
Let’s examine our training dataset for normality and skewness to see how the data is balanced.
trainBeers$flag = as.factor(trainBeers$flag)
trainBeers %>% select(abv, ibu, flag) %>% ggpairs(aes(color = flag)) + ggtitle("Distribution of Ale and IPA IBU data")
Below is the initial KNN training model. You will see that the model does a pretty good job at classifying between the two (definitely better than a straight guess). Additionally, it appears to perform specificity and sensitivity are very similar suggesting that the model is likely very balanced in it’s estimations.
##Classification Method 1: KNN
#Train the model (k=5)
classifications = knn(trainBeers[c(1,2)],testBeers[c(1,2)],trainBeers$flag, prob = TRUE, k = 5)
#The resulting confusion matrix
table(testBeers[,'flag'],classifications)
## classifications
## ALE IPA
## ALE 156 16
## IPA 24 91
CM_k5 = confusionMatrix(table(testBeers[,'flag'],classifications))
CM_k5
## Confusion Matrix and Statistics
##
## classifications
## ALE IPA
## ALE 156 16
## IPA 24 91
##
## Accuracy : 0.8606
## 95% CI : (0.8151, 0.8985)
## No Information Rate : 0.6272
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7064
##
## Mcnemar's Test P-Value : 0.2684
##
## Sensitivity : 0.8667
## Specificity : 0.8505
## Pos Pred Value : 0.9070
## Neg Pred Value : 0.7913
## Prevalence : 0.6272
## Detection Rate : 0.5436
## Detection Prevalence : 0.5993
## Balanced Accuracy : 0.8586
##
## 'Positive' Class : ALE
##
The above is based on k-value is five…what about k is equal to ten? You can see here that the accuracy is the same, with the increased specificity offsetting a decrease in sensitivity.
#Alternative K to train the model (k=10)
classifications_k10 = knn(trainBeers[c(1,2)],testBeers[c(1,2)],trainBeers$flag, prob = TRUE, k = 10)
#The resulting confusion matrix
table(testBeers[,'flag'],classifications_k10)
## classifications_k10
## ALE IPA
## ALE 156 16
## IPA 27 88
CM_k10 = confusionMatrix(table(testBeers[,'flag'],classifications_k10))
CM_k10
## Confusion Matrix and Statistics
##
## classifications_k10
## ALE IPA
## ALE 156 16
## IPA 27 88
##
## Accuracy : 0.8502
## 95% CI : (0.8035, 0.8894)
## No Information Rate : 0.6376
## P-Value [Acc > NIR] : 9.659e-16
##
## Kappa : 0.683
##
## Mcnemar's Test P-Value : 0.1273
##
## Sensitivity : 0.8525
## Specificity : 0.8462
## Pos Pred Value : 0.9070
## Neg Pred Value : 0.7652
## Prevalence : 0.6376
## Detection Rate : 0.5436
## Detection Prevalence : 0.5993
## Balanced Accuracy : 0.8493
##
## 'Positive' Class : ALE
##
In order to properly “tune” our hyperparameter, k, we will perform an interative loop with 100 random samples and 90 different k-values. From there, we calculate the average model accuracy in order to choose the best k-value.
##################################################################
# Loop for many k and the average of many training / test partition
set.seed(1)
iterations = 100
numks = 90
splitPerc = .7
masterAcc = matrix(nrow = iterations, ncol = numks)
for(j in 1:iterations)
{
trainIndices = sample(1:dim(test_4KNN)[1],round(splitPerc * dim(test_4KNN)[1]))
train = test_4KNN[trainIndices,]
test = test_4KNN[-trainIndices,]
for(i in 1:numks)
{
classifications = knn(train[,c(1,2)],test[,c(1,2)],train$flag, prob = TRUE, k = i)
table(classifications,test$flag)
CM = confusionMatrix(table(classifications,test$flag))
masterAcc[j,i] = CM$overall[1]
}
}
Based on the loop above, the following plot can be generated from the different k-values based on the 100 different random samples.
MeanAcc = colMeans(masterAcc)
plot(seq(1,numks,1),MeanAcc, type = "l")
Based on the data plotted above, the optimal K-value is listed below, along with the maximum accuracy of the model.
which.max(MeanAcc)
## [1] 5
max(MeanAcc)
## [1] 0.8569751
After creating our IPA estimation model, we will now do some additional exploratory analysis in order to find the best market to introduce a new Budweiser IPA beer. Initially, we will use the grep function to tally only IPA values and then filter our dataset accordingly.
# Part 9: Additional Data Exploration
# Plan - to count % IPA per total craft beers in each state.
# finding the beers containing "IPA" in their style and tallying
cat_freq <- main %>% mutate(category=
case_when(
grepl("IPA", beer_style) ~ "IPA",
TRUE ~ "OTHER"
)
) %>% group_by(state, category) %>% summarise(n=n()) %>% mutate(freq = n/sum(n)) %>% filter(category=="IPA")
cat_freq$state <- trimws(cat_freq$state)
# adding full state names from merging with us_map() library function
cat_freq <- merge(distinct(us_map(), full, abbr), cat_freq, by.x="abbr", by.y="state", all.x=TRUE) %>% mutate(n=replace_na(n, 0), freq=replace_na(freq, 0))
Next, let’s calculate our relative frequency of IPA to total craft beers per state.
# create percentage from frequency
cat_freq$freq = round(cat_freq$freq, digits=4)*100
After realizing the impact of population on market potential, I downloaded the 2019 consensus data in order to perform a “correction” for state population. My goal was to the best balance of IPA count per 100k consumers in order to find the best markets to perform trials before scaling to mass production.
# creatig dataframe from state consensus
state_pop <- data.frame(read.csv(here("project_files", "2019 consensus data.csv")))
# merging dataframe with %IPA per state dataframe
normalized_IPA <- merge(state_pop, cat_freq, by.x="State", by.y="full")
# normalizing IPA count per 100k state population
normalized_IPA <- normalized_IPA %>% mutate(IPA_per_100k = n/(Pop / 100000)) %>% arrange(desc(IPA_per_100k))
# creating a hover tag for the map
normalized_IPA$hover <- with(normalized_IPA, paste(State, '<br>', "IPA Beers", n, "<br>","IPA % of total craft beers", freq,"%", "<br>", "Population Rank", rank, "<br>", "IPA per 100k", IPA_per_100k))
Based on this evidence, you’ll see the following states are the best candidates for a trial run of Budeweiser IPA, according to population and IPA interest.
# displaying a table of top 5 markets for reference
fig8 <- normalized_IPA %>% select(State, Pop, IPA_per_100k) %>% top_n(8)
## Selecting by IPA_per_100k
fig8 %>% plot_ly(type='table',
header = list(
values = c("State", sprintf("State Population (%s Sample Size)",sum(fig8$Pop)), "IPA per 100k People")),
cells=list(
values=t(unname(as.matrix(fig8))),
font = list(color = c('#506784'), size = 12)
))
fig8
## State Pop IPA_per_100k
## 1 Vermont 627180 1.9133263
## 2 Colorado 5770545 0.9877750
## 3 Alaska 735720 0.9514489
## 4 Montana 1074532 0.7445102
## 5 Oregon 4245901 0.7301159
## 6 Indiana 6718616 0.5060566
## 7 Idaho 1790182 0.4468819
## 8 Maine 1342097 0.3725513
Now that we have normalized our dataset, let’s identify the states with the best potential for the IPA release. You can see from the results below, that if we were to pick a region for the release instead of isolating to only the top five consumer states, it appears that the Pacific Northwest to Colorado may be the best market to enter with hard initial avoidance in the Southeast Region.
# displaying all US IPA markets on map
fig9 <- plot_geo(normalized_IPA, locationmode = 'USA-states') %>%
add_trace(
z = ~IPA_per_100k, text=~hover, locations = ~abbr,
color = ~IPA_per_100k,
marker = list(line = list(color = toRGB(bg_color), width = 2.25)),colorscale='MAGMA'
) %>%
colorbar(title = "IPA") %>%
layout(
title = 'IPA count per 100k',
font = list(color = 'white'),
geo=list(
scope = 'usa',
projection = list(type = 'albers usa'),
showlakes = FALSE,
bgcolor=toRGB(bg_color, alpha = 1)),
paper_bgcolor=toRGB(bg_color, alpha = 1),
margin=list(l=20, r=20, t=60, b=20)
)
fig9